nemo_rl.data.datasets.preference_datasets.mmpr#

MMPR preference data for offline image MPO.

Module Contents#

Classes#

MMPRPreferenceDataset

Load the legacy MMPR meta-recipe as a canonical preference dataset.

Functions#

_path_signature

Return a cheap source signature suitable for cache invalidation.

format_mmpr_preference_dataset

Convert an MMPR row to the standard two-completion preference schema.

Data#

API#

nemo_rl.data.datasets.preference_datasets.mmpr._IMAGE_PLACEHOLDER_RE#

‘compile(…)’

nemo_rl.data.datasets.preference_datasets.mmpr._CACHE_VERSION#

2

nemo_rl.data.datasets.preference_datasets.mmpr._path_signature(path: pathlib.Path) → str#

Return a cheap source signature suitable for cache invalidation.

nemo_rl.data.datasets.preference_datasets.mmpr.format_mmpr_preference_dataset(
example: dict[str, Any],
) → dict[str, Any]#

Convert an MMPR row to the standard two-completion preference schema.

class nemo_rl.data.datasets.preference_datasets.mmpr.MMPRPreferenceDataset(
data_path: str,
split_validation_size: float | int = 0.01,
legacy_validation_split: bool = False,
seed: int = 42,
max_samples: int | None = None,
cache_dir: str | None = None,
**_: Any,
)#

Bases: nemo_rl.data.datasets.raw_dataset.RawDataset

Load the legacy MMPR meta-recipe as a canonical preference dataset.

Initialization